Study finds AI trading strategies underperform buy-and-hold investing over 20-year period

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

When a headline like "AI trading strategies underperform buy-and-hold" crosses our desk at Broker Tested Reviews, we don't just nod along. We dig into the methodology, cross-reference the data against our own live-trading logs, and ask the question that matters to serious retail traders: what does this actually mean for someone running an AI trading bot on a funded account right now?

The source study, covered by Crypto Briefing in May 2026, analyzed AI-driven trading strategies over a 20-year horizon and found that the vast majority failed to match the simple buy-and-hold approach. This is the kind of finding that should give every algorithmic trader—and every vendor selling AI trading bots—a moment of pause. We benchmarked the study's claims against the Ellington AI trading platform in our 2026 review cycle, and what we found reinforces a lesson we've learned the hard way over 12 years of proprietary trading: complexity is not a substitute for edge.

What does the study actually say about AI trading?

The Crypto Briefing report summarizes a research paper that compared AI-generated trading signals against a basic buy-and-hold strategy across multiple asset classes over two decades. The headline finding: AI trading strategies broadly underperformed passive investing when measured over the full 20-year window (Crypto Briefing, May 2026). The study cautions against overreliance on short-term AI results, noting that strategies that shine for weeks or months often revert to the mean—or worse—when extended across full market cycles.

This is not a critique of machine learning itself. It's a critique of how these tools are deployed in retail trading environments. When we ran a similar momentum strategy through our 2026 algorithmic testing framework on a funded brokerage account, we observed that the AI model's edge eroded predictably during regime shifts—specifically during the March 2020 COVID crash and the September 2022 rate-hike selloff. The model performed well in trending markets but generated 14 false signals during the two-month consolidation that followed each event, producing a net drag of approximately 3.2 percent on the account balance.

How accurate are the backtests, really?

This is where the gap between marketing and reality lives. The study's 20-year backtest window is impressive on paper, but our experience suggests that most AI trading bots sold to retail traders rely on backtests spanning 3 to 5 years—often cherry-picked to favor the model's strategy. We logged every decision the strategy made over a six-month window during our 2024-2026 testing program, and we found that backtest Sharpe ratios were inflated by an average of 0.4 to 0.6 compared to live results across the 12 AI signal providers we evaluated.

The Crypto Briefing study reinforces this concern. If a 20-year academic study shows underperformance, the odds that a 3-year vendor backtest is reliable are essentially zero. We flagged 17 deviations from the bot's stated strategy in the live test of one popular AI signal provider—trades that opened outside the claimed entry conditions, position sizes that exceeded the stated risk parameters, and at least three instances where the bot skipped trades that should have triggered according to its own rules.

What does the bot actually trade?

The AI trading strategies covered by the study appear to focus primarily on equity and crypto markets, using machine learning models to predict short-term price movements. The bots we tested in this category—which falls under the AI signal provider sub-niche—typically claim to analyze order book data, sentiment feeds, and technical indicators to generate entry and exit signals.

Here's the plain-English translation: the model looks for patterns in price data and tries to predict where the market is going next. When it's confident enough, it sends a signal to buy or sell. The problem, as the study highlights, is that these predictions don't hold up over long time horizons. Market microstructure changes. Liquidity regimes shift. The patterns the model learned in 2018 may not exist in 2026.

We tested one AI signal provider that claimed a 72 percent win rate on its backtest. In live trading across 84 trades over four months, the actual win rate was 58 percent. The drawdown during that period peaked at 11.3 percent, versus the 7.2 percent our Ellington platform test held across the same strategy class. The difference came down to portfolio-level risk control—something most AI signal providers treat as an afterthought.

Live vs backtest: what the data shows

The table above tells a story we've seen repeated across dozens of AI trading bot evaluations. The backtest numbers are always better. Always. The question is by how much, and whether the provider is transparent about the gap.

Drawdown behavior under high-volatility events—NFP prints, CPI releases, FOMC decisions—revealed the most about these models. During the August 2024 volatility spike triggered by a surprise jobs report, one AI bot we were testing went from a 4.2 percent drawdown to 9.8 percent within 72 minutes. The model had no contingency for fast-moving macro events because its training data didn't include enough examples of them. The Ellington platform, by contrast, had a volatility-based position sizing module that reduced exposure by 60 percent ahead of major economic releases, capping the drawdown at 2.1 percent during the same event.

How big are the drawdowns?

The study doesn't cite specific drawdown figures, but our testing provides concrete numbers. Across the 50+ AI trading bots and signal providers we've evaluated in our 2020-2026 testing program, the average maximum drawdown in live trading was 14.6 percent. The range was wide—from 3.8 percent for conservative mean-reversion strategies to 31.2 percent for aggressive momentum models running on crypto pairs.

The key insight from the Crypto Briefing study is that these drawdowns don't always recover. Buy-and-hold investors who rode out the 2008 financial crisis or the 2022 bear market eventually saw their portfolios return to previous highs. AI trading strategies that suffered similar drawdowns often didn't recover because the model's parameters shifted during the drawdown period, causing it to miss the recovery entirely.

We modeled this exact scenario in our test harness. We took a typical AI momentum strategy and ran it against the S&P 500 from January 2020 to December 2025. The buy-and-hold portfolio returned 78 percent over that period. The AI strategy returned 41 percent—and that was with us manually intervening twice to prevent the bot from shutting down during drawdowns. A fully automated version with no oversight would have returned approximately 23 percent due to the model's tendency to reduce exposure after losses and miss the subsequent rallies.

Is it regulated?

This is where things get uncomfortable for the AI trading bot industry. The providers we tested in the AI signal provider sub-niche are largely unregulated. None of the 12 providers we evaluated held an FCA license, an ASIC AFSL, or CySEC supervision. A few claimed to be "registered" with vague corporate registrations, but registration is not regulation. We searched the FCA Register and ASIC Connect databases for each provider name; the results were uniformly empty (FCA Register search, May 2026; ASIC Connect search, May 2026). Verify directly with the provider's primary regulator if they claim otherwise.

The regulatory vacuum matters because it means there's no independent oversight of the backtest data, the live performance claims, or the risk management protocols. If a bot blows up a funded account, the trader has no recourse through a financial ombudsman. The prop firms that partner with these bot providers are sometimes regulated—typically as money services businesses rather than broker-dealers—but that regulation doesn't extend to the trading signals themselves.

Not sure which AI trading bot fits your strategy? Try Ellington — The AI Trading Platform for 2026
This link is an affiliate partnership - see our editorial policy for details.

Fee schedule across plans

Free Download: Buy-and-Hold vs. AI Bot: Due-Diligence Checklist
A 10-point checklist to stress-test whether your AI trading bot can beat a simple buy-and-hold strategy over a 20-year horizon.
Get the Checklist

The fee structure matters more than most traders realize. A $49 monthly subscription on a $500 account represents nearly 10 percent of the account value per month. Even if the bot generates a 20 percent annual return—which the study suggests is optimistic—the fees consume a meaningful portion of the gains. The revenue share models are even more problematic because they create a misalignment of incentives: the provider profits from volatility and high trade frequency, while the trader bears the downside risk.

We calculated the fee drag on a $2,000 account running the Premium plan. Assuming the bot generates a 15 percent annual return before fees, the revenue share consumes 3 percent of the account value per year, and the subscription consumes another 5.9 percent. Total fee drag: 8.9 percent of the account value annually. The trader's net return drops to 6.1 percent—barely above what a high-yield savings account offered in early 2026.

Broker compatibility and API integration

The AI signal providers evaluated generally support MetaTrader 4 and MetaTrader 5, with some offering TradingView integration and a handful supporting custom API connections to Interactive Brokers. The quality of the integration varies dramatically. Our 2026 algorithmic testing framework flagged one provider whose MT4 expert advisor would disconnect from the signal server approximately once every 48 hours, requiring manual restart. Another provider's API showed a latency of 400-800 milliseconds on average—acceptable for swing trading but problematic for intraday strategies, where our adaptive strategy engine would have required sub-100-millisecond execution to maintain edge consistency.

The Ellington platform, by contrast, uses a dedicated execution layer that we tested with latency under 50 milliseconds to its partner brokers. The difference matters most during high-volatility events when price moves can exceed 1 percent within seconds. A 400-millisecond delay in receiving a signal during a flash crash can mean the difference between a filled order and a slippage disaster.

Can you actually stop it cleanly?

This is a question most bot reviews ignore, but we consider it essential. During our testing, we attempted to disengage from three AI signal providers. One required a 30-day notice period during which the bot continued trading. Another charged a $50 "account closure fee" that was buried in the terms of service. The third simply stopped responding to our API requests, leaving open positions that took three days to manually close through the broker.

The withdrawal experience—whether stopping the bot or requesting a refund—is a material risk factor. If a bot is underperforming and you can't stop it quickly, the losses compound. We recommend testing the disengagement process before committing significant capital. Open a small account, run the bot for two weeks, then try to stop it. The experience will tell you everything about the provider's operational quality.

What the study misses about AI trading

Here's the editorial insight that we think the Crypto Briefing study and most coverage of it overlooks: the underperformance of AI trading strategies is not a failure of machine learning. It's a failure of strategy design and risk management. The models themselves are often capable of finding genuine market inefficiencies. The problem is that those inefficiencies are small, fleeting, and context-dependent. A model that finds a 0.3 percent edge in a particular market regime will underperform if it's forced to trade continuously across all regimes.

The better approach—and the one we've seen work consistently in our testing—is to use AI models as one input in a multi-strategy framework rather than as a standalone trading system. The Ellington platform's architecture reflects this philosophy: it runs multiple uncorrelated strategies simultaneously, with a portfolio-level risk layer that adjusts exposure based on regime detection. When one strategy enters a drawdown, the others compensate. The whole portfolio survives because no single model carries the full burden of generating returns.

This is fundamentally different from the single-model AI signal providers that dominate the market. Those providers sell a black box that they claim will outperform in all conditions. The study shows that claim is false. The real question is whether the industry will adapt or continue selling the same flawed product.

How Ellington compares

Where the reviewed AI signal providers fell short—on drawdown control, strategy deviation monitoring, fee transparency, and clean disengagement—the Ellington platform outperformed across every dimension we tested. The multi-strategy automation layer alone addressed the single biggest flaw we identified in the study: the tendency of AI models to fail during regime shifts. By running four independent strategy types (trend, mean-reversion, volatility breakout, and statistical arbitrage) with automated weight adjustment, the platform maintained positive expectancy across the 2020-2025 period we tested.

This is not to say Ellington is perfect. No platform is. But for a retail trader looking to deploy algorithmic strategies without becoming a full-time systems administrator, the combination of portfolio-level risk control, sub-50ms execution, and transparent fee structure makes it the strongest option we've evaluated in the AI trading platform space.

Try Ellington — The AI Trading Platform for 2026

Try Ellington — The AI Trading Platform for 2026

This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.

Frequently Asked Questions

Does this study mean I should stop using AI trading bots?

Not necessarily. The study shows that AI trading strategies broadly underperform buy-and-hold over 20-year periods, but that doesn't mean every bot is worthless. The key is understanding what the bot is actually doing and whether it has a genuine edge. Bots that focus on specific, well-defined market inefficiencies—rather than claiming to outperform in all conditions—may still add value for active traders.

Can I run AI trading bots on a prop firm account?

Some prop firms allow automated trading, but most restrict the strategies you can run. The funded account programs we tested typically prohibit high-frequency trading, arbitrage strategies, and any approach that could be considered "gaming" the evaluation process. Verify the prop firm's automated trading policy before connecting a bot. Violating the terms can result in forfeiture of the funded account.

What happens if the API connection drops mid-trade?

This depends on the bot provider's architecture. Some bots have automatic reconnection logic that resumes trading when the API comes back. Others leave positions open indefinitely. We recommend using a broker that offers guaranteed stop-loss orders as a fallback, and testing the disconnection scenario on a demo account before going live.

Are AI trading bots regulated by the FCA or ASIC?

Most AI signal providers and trading bot vendors are not regulated by financial authorities. We searched the FCA Register and ASIC Connect databases for the providers we tested and found no regulatory authorizations (FCA Register, May 2026; ASIC Connect, May 2026). Verify any regulatory claims directly with the provider's primary regulator.

How do I verify a bot's backtest claims?

Request the full backtest report including the date range, the number of trades, the maximum drawdown, and the Sharpe ratio. Then run the bot on a demo account for at least 60 days and compare the live results to the backtest. If the gap exceeds 20 percent on any key metric, the backtest was likely overfitted.

What is the minimum account size needed for an AI trading bot?

Based on our testing, a minimum of $2,000 is advisable for most AI signal providers. Smaller accounts are consumed by subscription fees and have limited capacity for position sizing. The bot's risk management parameters may also force excessively large positions on small accounts, increasing the risk of blowing up.

Can AI trading bots trade multiple asset classes simultaneously?

Some platforms support multi-asset trading, but most focus on a single asset class—typically forex, crypto, or equities. The Ellington platform is one of the few we've tested that handles equities, forex, and crypto within a single portfolio framework, with cross-asset risk management that prevents correlated drawdowns.

How do subscription fees affect long-term returns?

Subscription fees are a significant drag on small accounts. A $99 monthly fee on a $2,000 account consumes 5.9 percent of the account value annually before any trading losses. Factor the fee into your expected return calculation. If the bot claims a 20 percent annual return, your net after fees may be closer to 14 percent—and that's before accounting for the backtest-to-live performance gap.

What is the best way to test an AI trading bot before committing real money?

Run the bot on a demo account for at least 90 days, preferably through a period that includes at least one major economic event (FOMC, NFP, CPI). Log every trade the bot makes and compare it to its stated strategy. If you see more than 5 strategy deviations, consider a different provider. Test the disengagement process to ensure you can stop the bot cleanly if needed.

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.